Validating Bounding Box

Uber's video-labeling system flags suspicious tracking annotations for people to review. These annotations are rectangles marking an object's position across frames. A rectangle can suddenly jump or switch to another object, but genuine movement and camera motion make a simple distance threshold unreliable.

The system examines appearance, motion and coordinates across an 11-frame window. A classifier trained on deliberately introduced annotation errors scores each frame, then groups suspicious frames into reviewable regions. When annotators submit their work, they can correct the flagged regions or dismiss mistaken suggestions. The described deployment focuses on position jumps and identity swaps, concentrating human review on likely problems. Architecture and review workflow.